Re: [PATCH net-next 00/12] ipvs: Add icmp scheduling

To: Patrick Schaaf <netdev@xxxxxx>
Subject: Re: [PATCH net-next 00/12] ipvs: Add icmp scheduling
Cc: Alex Gartrell <agartrell@xxxxxx>, Simon Horman <horms@xxxxxxxxxxxx>, Julian Anastasov <ja@xxxxxx>, kernel-team <kernel-team@xxxxxx>, lvs-devel@xxxxxxxxxxxxxxx
From: Alex Gartrell <alexgartrell@xxxxxxxxx>
Date: Thu, 20 Aug 2015 00:31:58 -0700
On Wed, Aug 19, 2015 at 11:09 PM, Patrick Schaaf <netdev@xxxxxx> wrote:
> Am 12.08.2015 22:47 schrieb "Alex Gartrell" <agartrell@xxxxxx>:
> Sounds like a missing feature in whatever does that ECMP thing, not being
> able to recognize ICMP / look at the ICMP inner L3/L4 addresses to drive the
> hashing. If you could get that to change, no partially-working (needing
> static schedule) IPVS hack would be neccessary, right?

If we were able to get the router to do the right thing, it would
definitely deal with the steady state case.  There are still two
problems though:
1) Doing this in the routers (in hardware) is /really/ hard.  Your
ASIC has to be able to descend into the icmp payload, which is
non-trivial.  For an example of where this routinely fails, simply
look at how routers and NICs handle ipv6 extension headers -- the
major NIC vendors simply give up immediately and omit the ports from
the hash.
2) It doesn't address the situation in which you want to swap or
restart and IPVS instance (as for sloppy TCP).  This will happen if
you add or remove next-hops via BGP or if you simply swap one.

tl;dr; it's too hard to change routers and, even if we did, it doesn't
address the "sloppy" scheduling case.

Alex Gartrell <agartrell@xxxxxx>
To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at

<Prev in Thread] Current Thread [Next in Thread>